LongJin-lab/ZeroSNet

Official Pytorch implementation of paper "Zero Stability Well Predicts Performance of Convolutional Neural Networks"

Zero Stability Well Predicts Performance of Convolutional Neural Networks

Official Pytorch implementation of paper "Zero Stability Well Predicts Performance of Convolutional Neural Networks" by Liangming Chen, Long Jin, and Mingsheng Shang.

Examples

Train ZeroSNet44_Opt on CIFAR-10:

CUDA_VISIBLE_DEVICES=0 python train_cifar_ZeroSNet.py --arch ZeroSNet44_Opt --dataset cifar10

Train a third-order-discretization CNN with coefficients [1, 1, 1, 1] on CIFAR-10 (Note that these coefficients lead to a non-zero-stable CNN):

CUDA_VISIBLE_DEVICES=0 python train_cifar_ZeroSNet.py --arch ZeroSNet44_Opt --dataset cifar10 --given_coe 1 1 1 1

Train ZeroSNet56_Tra on CIFAR-100:

CUDA_VISIBLE_DEVICES=0 python train_cifar_ZeroSNet.py --arch ZeroSNet56_Tra --dataset cifar100

Train ZeroSNet_Opt on ImageNet:

CUDA_VISIBLE_DEVICES=7,6,5,4,3,2,1,0 python3 -m torch.distributed.launch --nproc_per_node=8 --master_port 12345 main_ZeroSNet_IN.py --arch zerosnet18_in -bs 128 --lr 0.2 --opt_level O2 --data <your data set path>  --workers 8 --given_coe 0.3333333 0.5555556 0.1111111 1.77777778 

Evaluate robustness a third-order-discretization CNN with coefficients [1, 1, 1, 1] on CIFAR-10:

CUDA_VISIBLE_DEVICES=0 nohup python3 robustness_eval.py --arch + ZeroSNet56_Opt --noise_type rand --noise_coff 0.1 --dataset cifar10 --resume True --given_ks 1 1 1 1 --save_path <your save path> --workers 4